U.S. patent number 7,043,420 [Application Number 09/734,154] was granted by the patent office on 2006-05-09 for trainable dynamic phrase reordering for natural language generation in conversational systems.
This patent grant is currently assigned to International Business Machines Corporation. Invention is credited to Adwait Ratnaparkhi.
United States Patent |
7,043,420 |
Ratnaparkhi |
May 9, 2006 |
Trainable dynamic phrase reordering for natural language generation
in conversational systems
Abstract
A system and method to facilitate natural language generation in
a human-to-machine conversational system that produces written or
spoken output. In one aspect, a user provides a scoring function
and grammar rules including words and attributes. A method
according to the present invention then generates possible
reorderings of the words and attributes using the grammar rules and
determines an optimal ordering of the words and attributes using
the scoring function, which is then returned to the user.
Inventors: |
Ratnaparkhi; Adwait (Monroe,
NY) |
Assignee: |
International Business Machines
Corporation (Armonk, NY)
|
Family
ID: |
24950530 |
Appl.
No.: |
09/734,154 |
Filed: |
December 11, 2000 |
Prior Publication Data
|
|
|
|
Document
Identifier |
Publication Date |
|
US 20020116173 A1 |
Aug 22, 2002 |
|
Current U.S.
Class: |
704/9; 704/258;
704/260 |
Current CPC
Class: |
G06F
40/56 (20200101) |
Current International
Class: |
G06F
17/27 (20060101); G10L 13/00 (20060101) |
Field of
Search: |
;704/9,7,10,258,260 |
References Cited
[Referenced By]
U.S. Patent Documents
Other References
Dale and Reiter Tutorial on Building Natural Language Generation
Systems, EACL-99. cited by examiner .
Asahara et al. Jul. 2003, Extended Models and Tools for
High-performance Part-of-speech Tagger, Proceedings of the 17th
conference on Computational linguistics, vol. 1, 21-27. cited by
examiner .
Reiter et al., 1995, Building Applied Natural Language Generation
Systems, Cambridge University Press. cited by examiner .
Wang Weiwei et al., 1996, A natural Language Generation System
Based on Dynamic Knowledge Base, Proceeding of ICSP. cited by
examiner .
Berger et al., 1996, , A Maximum Entropy Approach to Natural
Language Processing, Association for Computational Linguistics.
cited by examiner .
Patten et al., Oct. 1991, Real-Time Generation of Natural Language,
IEEE Intelligent Systems, vol.: 6, Issue: 5, pp.: 15-22. cited by
examiner .
Langkilde, Irene, Generation that Exploits Corpus-Based Statistical
Knowledge, 1998, Proceedings of the 36th conference on Association
for Computational Linguistics--vol. 1, pp. 704-710. cited by
examiner .
Pieraccini et al. A Learning Approach to Natural Language
Understanding. Jun. 1, 1994, proceedings of the 1993 NATO ASI
Summer School, pp. 1-19. cited by examiner.
|
Primary Examiner: Chawan; Vijay
Assistant Examiner: Shortledge; Thomas E.
Attorney, Agent or Firm: F.Chau & Associates, LLC
Government Interests
The U.S. Government has a paid-up license in this invention and the
right in limited circumstances to require the patent owner to
license others on reasonable terms as provided for by the terms of
Government Contract Number MDA972-97-C-0012.
Claims
What is claimed is:
1. A computer-based method of generating natural language,
comprising the steps of: receiving a concept comprising attributes
and corresponding values of each of said attributes from a user;
receiving grammar rules from the user, each rule including a head,
a phrase fragment, a direction and a condition, wherein each phrase
fragment includes one of said attributes; receiving a scoring
function from the user; generating possible natural language
phrases using the grammar rules; determining an optimal natural
language phrase from the possible natural language phrases using
the scoring function; and returning said optimal natural language
phrase to the user.
2. The method of claim 1, wherein the head is a word.
3. The method of claim 1, wherein the phrase fragment is a natural
language phrase fragment.
4. The method of claim 1, wherein the direction indicates a
location of the phrase fragment.
5. The method of claim 1, wherein the condition is a code fragment
for restricting use of a rule.
6. The method of claim 1, wherein each attribute in the optimal
natural language phrase is replaced with its corresponding
value.
7. The method of claim 1, wherein the optimal natural language
phrase is a highest scoring natural language phrase that is
consistent with the grammar rules.
8. The method of claim 1, wherein the scoring function is a
probability P of a phrase of length N based on the probability P of
a word w.sub.i conditioned on two previous words, w.sub.i-1 and
w.sub.i-2, the scoring function comprises the equation:
.PI..sub.i=1 . . . NP(w.sub.i|w.sub.i-1,w.sub.i-2).
9. The method of claim 1, wherein the attributes are variables.
10. The method of claim 4, wherein the direction indicates that the
location of the phrase fragment is right of the head.
11. The method of claim 4, wherein the direction indicates that the
location of the phrase fragment is left of the head.
12. A program storage device readable by a machine, tangibly
embodying a program of instructions executable by the machine to
perform the method steps for generating natural language, the
method comprising the steps of: receiving a concept comprising
attributes and corresponding values of each of said attributes from
a user; receiving grammar rules from the user, each rule including
a head, a phrase fragment, a direction and a condition, wherein
each phrase fragment includes one of said attributes; receiving a
scoring function from the user; generating possible natural
language phrases using the grammar rules; computing a score for
each of the natural language phrases using the scoring function;
returning the highest scoring natural language phrase to the
user.
13. The program storage device of claim 12, wherein the head is a
word.
14. The program storage device of claim 12, wherein the phrase
fragment is a natural language phrase fragment.
15. The program storage device of claim 12, wherein the direction
indicates a location of the phrase fragment.
16. The program storage device of claim 12, wherein the condition
is a code fragment for restricting use of a rule.
17. The program storage device of claim 12, wherein each attribute
in the optimal natural language phrase is replaced with its
corresponding value.
18. The program storage device of claim 12, wherein the optimal
natural language phrase is a highest scoring natural language
phrase that is consistent with the grammar rules.
19. The program storage device of claim 12, wherein the scoring
function is a probability P of a phrase of length N based on the
probability P of a word w.sub.i conditioned on two previous words,
w.sub.i-1 and w.sub.i-2, the scoring function comprises the
equation: .PI..sub.i=1 . . . NP(w.sub.i|w.sub.i-1,w.sub.i-2).
20. The program storage device of claim 12, wherein the attributes
are variables.
21. A computer-based method of generating natural language,
comprising the steps of: receiving a concept comprising attributes
and corresponding values of each of said attributes from a user;
receiving grammar rules from the user, each rule including a head,
a phrase fragment, a direction and a condition, wherein each phrase
fragment includes one of said attributes; receiving a scoring
function from the user; generating a plurality of possible natural
language phrases using the grammar rules; determining an optimal
natural language phrase from the plurality of possible natural
language phrases using the scoring function; and returning said
optimal natural language phrase to the user; wherein the head is a
word; wherein the phrase fragment is a natural language phrase
fragment comprising a plurality of words; wherein the direction
indicates a location of the phrase fragment; wherein the condition
is a code fragment for restricting use of a rule using a binary
evaluation during run-time.
Description
BACKGROUND
1. Technical Field
The present invention relates to conversational systems, and more
specifically to a system and method to facilitate natural language
generation in a human-to-machine conversational system that
produces written or spoken output.
2. Discussion of the Related Art
Human beings communicate ideas with one another using a mechanism
known as natural language. Natural language evolved as a medium of
communication as human beings learned to communicate with one
another. However, due to the inherent structure of natural
language, it is an imperfect mechanism for conveying ideas. The
human brain translates natural language into concepts and ideas,
and allows communication between different individuals using
natural language through a complicated translation process that no
machine has been able to accurately duplicate.
A computer can generate written or spoken language output, but the
structure of the language from a computer rarely resembles natural
human language. Typically, prior art computer generated speech
stores a limited number of sentences which can be expressed at
predetermined times and in predetermined ways, which limits the
expressiveness of computer-generated language.
For example, in a conversation with a conversational system, a user
supplies the system with information in the form of a statement or
a question. The system then responds to the user with a statement
or a question. This exchange continues until the computer fulfills
the user's request.
The information in a simple conversation can be represented with
pairs of attributes and values. An attribute A and its
corresponding value V are written in the form {A=V}. For example,
the statement "a flight leaves at 3PM" in the domain or realm of
air travel can be represented with the attribute-value pair
{$timeDep="3PM"}, where $timeDep is the attribute denoting the
departure time, and "3PM" is the textual instantiation of the
attribute.
The majority of current conversational systems perform natural
language generation (NLG) with templates. Templates comprise
attributes interspersed between words of natural language. When the
system requires a phrase, it first chooses a template, and then
replaces the attributes in the template with their corresponding
values in the run-time environment. For example, the template "a
flight leaves at $timeDep" would be expanded to "a flight leaves at
3 PM" if the run-time environment contained the attribute-value
pair {$timeDep="3 PM"}. Given a set of attribute-value pairs, a
template provides a fixed way of rendering them into natural
language. However, using such a natural language generation method
with templates requires that a programmer write a different
template for every new phrase to be created.
Accordingly, an accurate and dynamic technique for automatically
and efficiently generating natural language is highly
desirable.
SUMMARY OF THE INVENTION
The present invention is directed to a system and method for
generating natural language by automatically determining possible
reorderings of words and attributes, and then determining an
optimal ordering of the words and attributes.
In an aspect of the present invention, a method for computer-based
generation of natural language is provided comprising the steps of
receiving a concept comprising attributes and corresponding values
of each of said attributes from a user, receiving grammar rules
from the user, each rule including a head, a phrase fragment, a
direction and a condition, wherein each phrase fragment includes
one of said attributes, receiving a scoring function from the user,
generating possible natural language phrases using the grammar
rules, determining an optimal natural language phrase from the
possible natural language phrases using the scoring function, and
returning said optimal natural language phrase to the user.
These, and other aspects, features, and advantages of the present
invention will be described or become apparent from the following
detailed description of preferred embodiments, which is to be read
in connection with the accompanying drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is an exemplary flow chart depicting a method of data flow
in natural language generation.
FIG. 2 depicts an exemplary grammar comprising two grammar rules
for generating flight descriptions in an air travel domain.
FIG. 3 depicts an exemplary graph algorithm showing two possible
complete phrases and their respective intermediate phrases that may
result from the two grammar rules in FIG. 2.
DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS
It is to be understood that the exemplary system modules and method
steps described herein may be implemented in various forms of
hardware, software, firmware, special purpose processors, or a
combination thereof. Preferably, the present invention is
implemented in software as an application program tangibly embodied
on one or more program storage devices. The application program may
be executed by any machine, device or platform comprising suitable
architecture. It is to be further understood that, because some of
the constituent system modules and method steps depicted in the
accompanying Figures are preferably implemented in software, the
actual connections between the system components (or the process
steps) may differ depending upon the manner in which the present
invention is programmed. Given the teachings herein, one of
ordinary skill in the related art will be able to contemplate these
and similar implementations or configurations of the present
invention.
A natural language generation method according to the present
invention dynamically determines word and attribute order based on
one or more aspects of the dialog state. The method chooses between
many possible grammatical phrases for a given set of
attribute-value pairs to find the phrase with the most appropriate
word and attribute ordering.
The ability to reorder words and attributes is important for the
purposes of generating natural language. For example, in a
conversation concerning the domain of air travel where information
on departure times of flights out of Boston is being sought, the
phrase "A flight from Boston that leaves at 3 PM" might be a more
appropriate response that is more akin to what would be said in
natural language than the phrase "A 3PM flight that leaves from
Boston" even though it expresses the same information. The present
invention is a domain-independent software framework that can
efficiently and automatically perform dynamic reordering of words
and attributes such that a most appropriate phrase for a given
situation is found, and that does not require the programmer to
specify all the possible orderings.
FIG. 1 is an example of a flow chart depicting a method of data
flow in natural language generation in which a semantic
representation of a concept 101 is taken as input by a natural
language generation (NLG) module 103, and a phrase in natural
language 105 that corresponds to the concept 101 is produced as
output. A concept comprises attribute-value pairs. For example, a
semantic representation of a concept in the air travel domain may
be: {$locFr="New York", $locTo="Boston"} Possible corresponding
English natural language phrases may be: "A flight from New York to
Boston" "A flight from New York that arrives in Boston" "A flight
to Boston that leaves from New York"
The NLG module 103 requires grammar rules and a scoring function to
be provided in advance.
The grammar rules preferably include a head, a direction, a phrase
fragment and a condition. For exemplary purposes, the head is a
user-specified word, which is typically the main word of the
phrase, and the direction indicates the location of a phrase
fragment as being, for example, left or right of the head. The
phrase fragment is preferably a natural language phrase, and the
condition is preferably a code fragment that evaluates to either 1
(true) or 0 (false) in the run-time environment of the
conversational system. Any domain-specific restrictions on the
usage of a particular grammar rule, for example, can be encoded in
the condition of the grammar rule.
The scoring function comprises a predetermined formula to compute
the score of a generated phrase. Preferably, N-gram language models
that are used in known speech recognizers, are used, which will be
described in further detail below.
A system according to the present invention finds an optimal
natural language phrase by first using the provided grammar rules
to generate a set of possible natural language phrases, and then
searching this set of possible phrases for a highest scoring
natural language phrase according to the predetermined scoring
function.
Before the grammar rules can be applied, the condition of the
grammar must be satisfied with respect to the phrase fragment to
which it will be applied. A natural language phrase for describing
the input semantic representation of a concept can then be
constructed by several applications of grammar rules to the
user-specified word (or head) which is typically the main word of
the phrase. The particular sequence in which grammar rules are
applied ultimately determines the word order within a generated
natural language phrase. For example, when a direction is specified
as "right", the phrase fragment of the rule is added to the right
of the head and preferably to the right of any previously added
phrase fragments to the head when the condition is satisfied.
Similarly, when a direction is specified as "left", the phrase of
the rule is added to the left of the head and preferably to the
left of any previously added phrases to the head when the condition
of the rule is satisfied.
FIG. 2 depicts two exemplary grammar rules for generating flight
descriptions in an air travel domain. For illustrative purposes,
the head 203 is specified as being the word "flights". Rule 1 (201)
is interpreted to read that under a condition 209 that an attribute
"$locTo" has not yet been generated, than a phrase fragment 207 "to
$locTo" should occur somewhere in a direction 205 that is to the
right of the head "flights". The interpretation of rule 2 (250) is
that if a condition 209 is satisfied in which a "$locFr" attribute
has not yet been generated, a phrase fragment 207 "from $locFr"
should occur somewhere in a direction that is to the right of the
head "flights".
It is readily apparant to one skilled in the art that the
conditions (or code fragments) (209) are normally implemented as
fragments of an interpreted language such as, for example, TCL,
C++, etc. Preferably, the illustrative conditions 209 provide that
each attribute is only generated once to prevent a phrase
describing the same attribute from being produced and inserted into
the same phrase ordering. This is shown here, for example, in Rule
1 (201) which has a condition 209 specifying that the phrase 207
"to $locTo" should only be used if the attribute $locTo has not yet
been generated. Similarly, the condition in Rule 2 (250) prevents
the use of the phrase "from $locFr" unless the attribute "$locFr"
has not yet been generated.
Although the programmer specifies grammar rules which include
various words and phrase fragments, the orderings of these words
and phrase fragments relative to one another are not specified.
Instead, a system according to the present invention automatically
generates a set of possible phrase orderings using the grammar
rules and then searches this set for an "optimal" phrase ordering
which includes all of the attributes in the semantic
representation. An optimal phrase ordering is consistent with the
grammar rules specified by a user and is a highest scoring phrase
according to a predetermined scoring formula. The attributes $locTo
and $locFr will be substituted with their respective natural
language values at the end of the generation process when the
system finds the optimal phrase ordering.
In general, there may be hundreds or thousands of possible phrase
orderings for a complex grammar rule. Advantageously, an aspect of
the present invention provides that the number of grammar rules a
programmer actually needs to write to generate such a large number
of possible orderings is quite small. For example, an N number of
rules written by the programmer can generate N factorial possible
phrase reorderings.
The scoring function can be preprogrammed to return for example,
the top 5, 10, etc. phrase orderings. In general, any scoring
function can be used, such as N-gram language models that are used
in known speech recognizers. Typically, a large sample database of
text in the domain or genre of interest is required by the system
to compute the score of a generated phrase.
The language model provides an estimate of the probability of a
word sequence W for the given phrase generation task. To
illustrate, the probability of a sequence of words is the
probability of each individual word conditioned on, for example,
the previous 2 words. The score of a phrase ordering can be given
by the formula: .PI..sub.i=1 . . . NP(w.sub.i|w.sub.i-1,w.sub.i-2)
where P(w.sub.i|w.sub.i-1, w.sub.i-2) is a conditional probability
model, N is the length of the phrase, and w.sub.i is the word w in
position i of the phrase. The probability model is trained from the
word co-occurrence statistics observed in the large text
database.
In one aspect, a system according to the present invention uses a
graph algorithm to search a fraction of the often numerous possible
phrases that are generated by the grammar rules. In general, there
may be hundreds or thousands of possible phrases resulting from a
complex grammar function.
For illustrative purposes, FIG. 3 depicts a simplified example of a
graph algorithm showing two possible complete phrases and their
respective intermediate phrases that may result from the two
grammar rules in FIG. 2. Heads and phrase fragments are represented
as vertices in the graph. For example, the head "flights" is
represented as an original vertex 301, the phrase fragment "flights
from $locFr" is represented as a secondary vertex 303 and the
phrase "flights to $locTo" is indicated as a secondary vertex 305.
The vertices are connected by directed edges 302 which represent
grammar rules. A set of vertices that has no outgoing directed
edges is referred to as a frontier of the graph. In FIG. 3, the
frontier is represented by vertices 307 and 309.
The system starts with an initial vertex that contains a
user-specified word or head (typically the main word in a phrase),
for example, the initial vertex 301 "flights". At the iteration of
each search, the system applies grammar rules to the vertices in
the existing frontier, thereby creating new frontier. Specifically,
an application of a grammar rule to a phrase of an existing vertex
results in a new, larger phrase, which is represented by a new
vertex. For example, a directed edge 302 comprising Rule 2 (250)
can be applied to the initial vertex 301 to result in an
intermediate phrase or secondary vertex 303. Rule 2 provides
placing the phrase fragment "from $locFr" to the right of the head
"flights" so long as the attribute "$locFr" has not yet been
generated. This produces the intermediate phrase 303 "flights from
$locFr".
Next, a second directed edge 306 comprising Rule 1 (201) is applied
to the intermediate phrase 303. Rule 1 dictates that the phrase
fragment "to $locTo" be placed to the right of the head "flights".
If there are any previously added phrase fragments after the head,
then the phrase fragment "to $locTo" can be placed to the right of
the previously added phrase fragments as well. This results in a
complete phrase 307 "flights from $locFr to $locTo" which has both
the attributes "$locFr" and "$locTo".
Phrases that mention all of the attributes in the input semantic
representation of a concept are called complete phrases. The
system's ultimate goal in the search is to find the vertex in the
graph with the highest scoring complete phrase. The system can be
programmed to return for example, a top 5 (K=5) number of highest
scoring phrases. Phrases which contain multiple instances of the
same attribute are disallowed, and low-scoring phrases are
discarded by the system. The search terminates when at least one of
the vertices in the new frontier comprises a complete phrase.
The result of the graph search will be a phrase that contains
natural language interspersed with attributes, such as for example,
"flights from $locFr to $locTo". Once the search is completed and a
highest scoring complete phrase found, the attributes are replaced
with their corresponding values specified in the input semantic
representation, so that the final result comprises natural
language. In the above example, the final result would be "flights
from New York to Boston".
In a similar process, a complete phrase 309 can be produced by
first applying a directed edge 304 comprising Rule 1 (201) to the
initial vertex 301 to result in an intermediate phrase 305, and
then applying a directed edge 308 comprising Rule 2 (250) to the
intermediate phrase 305.
The method of expansion used in the present invention is based on
the idea of context-free grammars. In one aspect, FIG. 3
illustrates an example in which a phrase fragment of a rule is
placed after any previous phrases of preceding rules. In another
aspect, recursive expansion may be used in which words specified in
the phrase fragments of one rule can be recursively expanded by
other rules. The search procedure will recursively apply rules in
order to find complete phrases. To illustrate, given the following
rules (the conditions have been omitted for clarity):
TABLE-US-00001 head direction phrase condition 1) A right B C 2) B
right D E 3) E left F G
when the initial word given by the user is "A", it can be expanded
via Rule 1 to "A B C". This can be further expanded using Rule 2 to
"A B D E C". Using Rule 3, this can be expanded again to give "A B
D F G E C".
In spoken dialogue, people typically express old information, i.e.,
information which is already known, at the beginning of a sentence
and new information, i.e., information which is desired or being
focused on, at the end of a sentence. Thus, in a conversational
system, the ability to dynamically reorder phrases is necessary if
the system is to properly emphasize the new information.
The present invention can implement this ability to reorder phrases
by, for example, modifying the scoring function to reflect a
preference for certain attribute orderings. To illustrate, each
attribute can be assigned an individual priority by the programmer.
The system can then generate a sentence in a way to reflect the
priorities.
Another way the present invention can implement the ability to
reorder phrases is by modifying the grammar rules. For illustrative
purposes, the following rules can be used in the domain of hotels
for putting new information (the "focus") at the end of a phrase,
and the old information at the beginning. In the course of a
conversation, certain attributes will represent new information,
while other attributes will represent old information. The
programmer can write two rules for each attribute, such that one
rule expresses the attribute when it denotes old information and
the other rule expresses the attribute when it denotes new
information.
For example, the rules to define the grammar can have the following
general format: nlg_addrule hotelgram [head][direction][phrase
expansion][TCL code fragment, i.e., the rule condition]
The exemplary words "nlg_addrule" and "hotelgram" are to inform the
system that a rule is being entered.
The following rules can be used to define the grammar: 1) An
example of a rule for when there is something in focus is:
nlg_addrule hotelgram.-{there is a room that}{[llength [array name
focus]]>0} 2) An example of a rule for when there is nothing in
focus is: nlg_addrule hotelgram.-{there is a room}{[llength [array
names focus]]==0}
In both of the above examples 1 and 2, the period "." after
"hotelgram" indicates that the initial head word is a period. The
"-" sign indicates that the direction is to the left of the head
word. 3) An example of a rule for when a city is in focus is:
nlg_addrule hotelgram that+&{is located in $city}{[info exists
city]&& [info exists focus(city)]}
Here, the word "that" is the head, and the "+" sign indicates that
the phrase fragment "is located in $city" must occur to the right
of the head. The "&" is an extension code which instructs the
system to automatically put in commas and the word "and" as needed
in the generated sentence. The words "info exists city" and "info
exists focus (city)" are conditions which ask, respectively, "do we
still have to generate a city variable?" and "is it in focus?" Both
these conditions must be satisfied for a sentence to be generated
with a city in focus. 4) An example of a rule for when a city is
not in focus is: nlg_addrule hotelgram room+{in $city}{[info exists
city]&& ![info exists focus(city)]}
Here, the word "room" is the head, and the phrase fragment "in
$city" will occur to the right of the head. The conditions which
must be satisfied for a sentence to be generated with a city that
is not in focus are "do we still have to generate a city variable?"
and "is it not in focus?"
After the rules for defining the grammar have been created and
input into the system, commands for actually generating the text
can be implemented.
These commands to generate text can have the following general
format:
nlg.sub.--gen hotelgram model [head][attribute-value list]{ }{ }
The words "nlg gen" are a command to instruct the system to
generate text. The following commands can be used to generate the
actual natural language text: 1) For example, a command for
generating a sentence with a city in focus is:
TABLE-US-00002 A) nlg_eval hotelgram { set focus(city) 1 } B) puts
[ nlg_gen hotelgram model. {$city {New York} $price {$200}
$roomtype {double}} {} {}] C) nlg_eval hotelgram { unset
focus(city) 1 }
Here, the words "nlg_eval" are a command which sets and unsets the
focus, and the word "puts" is a print command. Line A indicates to
the system that a city is in focus, line B comprises a list of
attribute/value pairs, and line C unsets the focus on the city
after the sentence is generated to return the system to its
original state.
An exemplary result from the above grammar rules and commands in
which a city is in focus is:
TABLE-US-00003 Focus Sentence city there is a double room under
$200 that is located in New York.
Advantageously, the present invention eliminates the need for a
programmer to specify every possible ordering of various words and
phrase fragments. Instead, templates which a programmer might
ordinarily have to write manually are generated automatically using
the above algorithm. The programmer only has to specify
attribute-value pairs, desired grammar rules and a scoring
function, and a system according to the present invention will
piece together the specified attributes, words and phrase fragments
to find possible reorderings which are consistent with the grammar
rules. The system will then use the scoring function to look for a
highest scoring reordering.
Although illustrative embodiments of the present invention have
been described herein with reference to the accompanying drawings,
it is to be understood that the present invention is not limited to
those precise embodiments, and that various other changes and
modifications maybe affected therein by one skilled in the art
without departing from the scope or spirit of the invention. All
such changes and modifications are intended to be included within
the scope of the invention as defined by the appended claims.
* * * * *